Risk Reduction in Large Portfolios: Why Imposing The Wrong Constraints Helps

Risk Reduction in Large Portfolios:
Why Imposing the Wrong Constraints Helps

Ravi Jagannathan and Tongshu Ma
ABSTRACT
Green and Hollield (1992) argue that the presence of a dominant factor is why we
observe extreme negative weights in mean-variance-ecient portfolios constructed using
sample moments. In that case imposing no-shortsale constraints should hurt whereas
empirical evidence is often to the contrary. We reconcile this apparent contradiction.
We explain why constraining portfolio weights to be nonnegative can reduce the risk
in estimated optimal portfolios even when the constraints are wrong. Surprisingly,
with no-shortsale constraints in place, the sample covariance matrix performs as well
as covariance matrix estimates based on factor models, shrinkage estimators, and daily
data.
Jagannathan is from the Kellogg School of Management at Northwestern University and the National Bureau
of Economic Research and Ma is from the David Eccles School of Business at the University of Utah. We thank
Torben Andersen, Gopal Basak, Louis Chan, Gregory Connor, Kent Daniel, Bernard Dumas, Ludger Hentschel,
Philippe Henrotte, Ioulia Ioe, Olivier Ledoit, Ludan Liu, Andrew Lo, Jesper Lund, Robert McDonald, Steve Ross,
Jay Shanken and seminar participants at Copenhagen Business School, HEC, INSEAD, MIT, Norwegian School of
Management, Hong Kong University of Science and Technology, University of Rochester, University of Illinois at
Urbana-Champaign, University of Utah, and the 2002 AFA meeting in Atlanta, and especially the referees and the
editor, Rick Green, for helpful comments. We are responsible for any errors or omissions.
Markowitzs (1952, 1959) portfolio theory is one of the most important theoretical developments
in nance. Mean-variance ecient portfolios play an important role in this theory. Such portfolios
constructed using sample moments often involve taking large short positions in a number of assets
and large long positions in some other assets. Since negative portfolio weights (short positions)
are dicult to implement in practice, most investors impose the constraint that portfolio weights
should be nonnegative when constructing mean-variance ecient portfolios.
Green and Hollield (1992) argue that because a single factor dominates the covariance struc-
ture, it is dicult to dismiss the observed extreme negative and positive weights as entirely due to
the imprecise estimation of the inputs. They note that minimum variance portfolios can be con-
structed in two steps. First, naively diversify over the set of high beta stocks and the set of low beta
stocks separately. The resulting two portfolios will have very low residual risk and dierent betas.
Next, short the high beta portfolio and long the low beta portfolio to get rid of the systematic risk.
Clearly, the second step involves taking extreme long and short positions when the dispersion in
the assets betas is small. This is why minimum variance portfolios take large positive and negative
positions in the underlying assets.
If extreme negative weights in ecient portfolios arise due to the presence of a single dominant
factor then it would appear that imposing the nonnegativity constraints would lead to a loss
in eciency. However, empirical ndings suggest that imposing these constraints improves the
eciency of optimal portfolios constructed using sample moments.
1
In this paper we show that imposing the nonnegativity constraint on portfolio weights can help
even when Green and Hollield are right i.e., the true covariance matrix is such that ecient
portfolios involve taking large negative positions in a number of assets. We show that each of the
no-shortsales constraints is equivalent to reducing the sample covariances of the corresponding asset
with other assets by a certain amount. Stocks that have high covariances with other stocks tend to
receive negative portfolio weights. Hence, to the extent that high estimated covariances are more
likely to be caused by upward-biased estimation error, imposing the nonnegativity constraint can
reduce the sampling error. It follows from the theory of shrinkage estimators that imposing the
no-shortsales constraint can help even when the constraints do not hold in the population.
We also study the impact of upper bounds on portfolio weights since they are also commonly
imposed by practitioners. We show that each upper bound constraint is equivalent to increasing the
sample covariances of the corresponding asset with other assets by a certain amount. Since stocks
that have low covariances with other stocks tend to get extreme high portfolio weights, and these
extreme low estimated covariances are more likely to be caused by downward-biased estimation
error, this adjustment in estimated covariances could reduce sampling error and help the out-
of-sample performance of the optimal portfolios. Our empirical evidence suggests that imposing
upper bounds on portfolio weights does not lead to a signicant improvement in the out-of-sample
performance of minimum risk portfolios when no-shortsales restrictions are already in place.
The most common estimator of a covariance matrix is the corresponding sample covariance
matrix of historical returns. It has long been recognized that mean-variance ecient portfolios
constructed using the sample covariance matrix perform poorly out of sample. See Jobson and
1
Korkie (1980, 1981), Frost and Savarino (1986, 1988), Jorion (1986), Michaud (1989), Best and
Grauer (1991), and Black and Litterman (1992). The primary reason is that the large covariance
matrices encountered in practice require estimating too many parameters. For example, the covari-
ance matrix of the returns of 500 stocks has about 125,000 distinct parameters. However, monthly
returns for all traded stocks are only available for the past 900 months or so. This gives less than
four degrees of freedom per estimated parameter. Consequently the elements of the covariance
matrix are estimated very imprecisely.
Several solutions to this problem have been suggested. The rst is to impose more structure on
the covariance matrix to reduce the number of parameters that have to be estimated. This includes
factor models and constant correlation models. The second approach is to use shrinkage estimators,
shrinking the sample covariance matrix toward some target, such as the single-index model (Ledoit
(1996, 1999)). The third approach is to use data of higher frequency; for example, to use daily
data in place of monthly data. These three approaches are widely used both by practitioners and
academics.
For any covariance matrix estimator, we can impose the usual portfolio weight constraints.
According to Green and Hollield (1992), these constraints are likely to be wrong in population
and hence they introduce specication error. According to our analysis, these constraints can
reduce sampling error. Therefore the gain from imposing these constraints depends on the tradeo
between the reduction in sampling error and the increase in specication error. For covariance
matrix estimators that have large sampling error, such as the sample covariance matrix, imposing
these constraints is likely to be helpful. However, for the factor models and shrinkage estimators,
imposing such constraint is likely to hurt. This is what we nd empirically. Sampling errors are
likely to be less important when the assets we use to construct minimum variance portfolios are
themselves large portfolios. In that case imposing nonnegativity constraints should hurt. Again
this is what we nd.
In this paper, we will focus on the eect of portfolio weight constraints on the global minimum
variance portfolios. We do this because it has been reported in the literature that the global
minimum variance portfolio has as large an out-of-sample Sharpe Ratio as other ecient portfolios.
2
We will present some evidence on this point in Section III. We also examine the eect of portfolio
weight constraints on the performance of the minimum tracking error portfolios. We are interested
in minimum tracking error portfolios because it may be necessary in practice to construct portfolios
using a subset of all available stocks that have low transactions costs and high liquidity, in order
to track certain benchmark indices that contain assets that are not actively traded.
Our main empirical ndings are the following:
1. For factor models and shrinkage estimators, imposing the usual portfolio weight constraint re-
duces the portfolios eciency slightly. On the other hand, when the no-shortsales restriction
is imposed, minimum variance and minimum tracking error portfolios constructed using the
sample covariance matrix perform almost as well as those constructed using factor models,
shrinkage estimators, or daily returns.
3
2
2. When shortsales are allowed, minimum variance portfolios and minimum tracking error port-
folios constructed using daily return sample covariance matrix perform the best. When co-
variance matrices are estimated using daily data, corrections for microstructure eects that
have been suggested in the literature do not lead to superior performance. So far, this has
not been recognized in the literature.
3. Tangency portfolios, whether constrained or not, do not perform as well as the global mini-
mum variance portfolios in terms of out-of-sample Sharpe Ratio. This means that the esti-
mates of the mean returns are so noisy that simply imposing the portfolio weight constraint
is not enough, even though the constraints still have a shrinkage eect.
4. Monte Carlo simulations indicate that when nonnegativity constraints are in place, global
minimum variance portfolios constructed using the sample covariance matrix can perform
just as well as the covariance matrix estimators constructed by imposing the factor structure,
even when returns do have a dominant factor structure in the population.
5. In the simulations, imposing the nonnegativity constraint typically worsens the performance
when a single factor model is used to estimate the covariance matrix. On the other hand, in
the data, imposing nonnegativity constraints does not hurt much when using factor models.
This suggests that stock returns probably do not have a stable time invariant factor structure.
The rest of the paper is organized as follows. In Section I, we provide a theoretical analysis of the
shrinkage-like eect of imposing the no-shortsales restriction and upper bounds on portfolio weights
when constructing global minimum risk portfolios. In Section II, we use simulation to evaluate the
tradeo between specication error and sampling error. This tradeo clearly depends on the true
covariance structure of the assets, which is not observed. We therefore calibrate the covariance
structure in the simulation to that of the U.S. stocks. Using the simulation we provide some
guidance as to when the nonnegativity constraint becomes counterproductive. In Section III, we
empirically examine the eects of portfolio weight constraints using the out-of-sample performances
of the constrained and unconstrained optimal portfolios. We also briey discuss the role of portfolio
weight constraints in constructing tangency portfolios. We conclude the paper in Section IV. Proofs
and details about the construction of various covariance matrix estimates using daily returns are
collected in the appendices.
I. The Eect of Portfolio Weight Constraints
A. Some Theoretical Results
Given an estimated covariance matrix S, the global portfolio variance minimization problem
when portfolio weights are constrained to satisfy both a lower bound of zero and an an upper bound
3
of is given by:
min
S (1)
s.t.
i
= 1 (2)
i
0, i = 1, 2, , N. (3)
i
, i = 1, 2, , N. (4)
The Kuhn-Tucker conditions (necessary and sucient) are
j
S
i,j
j

i
+
i
=
0
0, i = 1, 2, , N. (5)
i
0, and
i
= 0 if
i
> 0, i = 1, 2, , N. (6)
i
0, and
i
= 0 if
i
< , i = 1, 2, , N. (7)
Here = (
1
, ,
N
)
are the Lagrange multipliers for the nonnegativity constraints (3), =

(
1
, ,
N
)
the multipliers for the constraints (4), and

0
is the multiplier for (2).
Denote a solution to the constrained portfolio variance minimization problem (1)-(4) as
++
(S).
Let 1 denote the column vector of ones. Then we have the following proposition.
Proposition 1: Let
S = S + (1
+1
) (1
+1
). (8)
Then

S is symmetric and positive semi-denite, and
++
(S) is one of its global minimum variance
portfolios.
All proofs are given in Appendix A.
This result shows that constructing a constrained global minimum variance portfolio from S
is equivalent to constructing a (unconstrained) minimum variance portfolio from

S = S + (1
+
1
) (1
+1
). Later we will interpret

S as a shrunk version of S, and argue that this shrinkage
can reduce sampling error.
In general, given a constrained optimal portfolio
++
(S), there are many covariance matrix
estimates that have
++
(S) as their (unconstrained) minimum variance portfolio. Is there anything
special about

S? We do have an answer to this question when returns are jointly normal and S is
the MLE of the population covariance matrix.
Let the N 1 return vector h
t
= (r
1t,
, r
2t
, , r
Nt
)
be iid normal N(, ). Then the MLE of

is S =

T
t=1
(h
t
h)(h
t
h)
/T. The likelihood function depends on both and , even though

we want to estimate only. To get rid of the dependence on , recall that for any estimate of the
covariance matrix, the MLE of the mean is always the sample mean (Morrison (1990)). With this
estimate of the mean, the log-likelihood (as a function of the covariance matrix alone) becomes
l() = CONST
T
2
ln ||
T
2
tr(S
1
). (9)
This can also be considered as the likelihood function of
1
, and is dened for nonsingular .
4
Now consider the constrained MLE of , subject to the constraint that the global minimum
variance portfolio constructed from satises the weight constraints (3)-(4). Let
i,j
denote the
(i, j)-th element of and
i,j
denote the (i, j)-th element of
1
; then the constraints are
i,j
0, i = 1, 2, , N. (10)
i,j

k,j
. (11)
So the constrained maximum likelihood (ML) problem is maximizing (9), subject to constraints
(10)-(11). We have the following proposition.
Proposition 2: Assume that returns are jointly iid normal N(, ). Let S be the unconstrained
MLE of .
1. Given S, let {
i
,
i
,
i
}
N
i=1
be a solution to the constrained portfolio variance minimization
problem (1) - (4), and construct

S according to (8). Assume

S is nonsingular. Then

S and
{
i
, (1 )
i
}
N
i=1
jointly satisfy the rst order conditions for the constrained ML problem.
2. Let

S and {
i
,
i
}
N
i=1
jointly satisfy the rst order conditions for the constrained ML problem.
For i = 1, , N, dene
i
=

j
S
i,j
/
S
k,l
, the normalized row sums of

S
1
. Then
{
i
,
i
/(1 ),
i
}
N
i=1
is a solution to the constrained portfolio variance minimization problem
(1) - (4), given S.
Roughly speaking, this proposition says that the

S constructed from the solution to the con-
strained global variance minimization problem is the ML estimator of the covariance matrix, subject
to the condition that the global minimum variance portfolio weights satisfy the nonnegativity and
upper bound constraints. So we could impose the constraints in the estimation stage instead of the
optimization stage and the result would be the same.
When only the nonnegativity constraint is imposed, the vector of Lagrange multipliers for the
upper bound will be zero. So

S = S(1
+1
), and we can simplify the statements in Proposition

2 in a straightforward way.
B. A Shrinkage Interpretation of the Eects of Portfolio Weight Constraints
Lets rst examine the eect of the nonnegativity constraint. Consider the unconstrained global
portfolio variance minimization. The rst order condition is
N
j=1
j
S
i,j
=
0
0, i = 1, 2, , N. (12)
The above condition says that at the optimum, stock is marginal contribution to the portfolio
variance is the same as stock js, for any i and j.
5
Suppose stock i tends to have higher covariances with other stocks; i.e., the ith row of S tends
to have larger elements than other rows. Then stock is marginal contribution to the portfolio
variance, 2
N
j=1
j
S
i,j
, will tend to be bigger than other stocks marginal contributions. Therefore,
to achieve optimality, we need to reduce stock is portfolio weight. Stock i may even have negative
weight if its covariances with other stocks are suciently high. Therefore, a stock tends to receive
a negative portfolio weight in the global minimum variance portfolio if it has higher variance and
higher covariances with other stocks. In a one-factor structure, these are the high-beta stocks.
With only the nonnegativity constraint, Proposition 1 implies that constructing the constrained
global minimum variance portfolio from S is equivalent to constructing the (unconstrained) global
minimum variance portfolio from

S = S (1
+ 1
). Notice that the eect of imposing the

constraint is that whenever the nonnegativity constraint is binding for stock i, its covariances with
other stocks, S
ij
, j = i, are reduced by
i
+
j
, a positive quantity, and its variance is reduced by
2
i
. We saw before that a stock will receive a negative portfolio weight in the minimum variance
portfolio if its covariances are relatively too high. Therefore, the new covariance matrix estimate,

S,
is constructed by shrinking the large covariances toward the average covariances. Since the largest
covariance estimates are more likely caused by upward-biased estimation error, this shrinking may
reduce the estimation error.
We can interpret the eect of the upper bounds on portfolio weights similarly. In the uncon-
strained portfolio variance minimization, those stocks with low covariances with other stocks tend
to receive high portfolio weights. So when the upper bound is imposed, it tends to be binding for
these stocks. Proposition 1 says that constructing the upper-bound constrained minimum variance
portfolio using S is the same as constructing the unconstrained minimum variance portfolio from
S = S + (1
+1
). So the eect of imposing the constraint is that, whenever the upper bound is
binding for stock i, its variance is raised by 2
i
and its covariance with another stock j is increased
by (
i
+
j
). Since the upper bound tends to be binding for those stocks with low covariances with
other stocks, and low estimated covariances are more likely to be plagued by downward-biased
estimation error, this adjustment may reduce sampling error and achieve better estimates of the
covariance matrix in certain sense. This adjustment has a shrinkage-like eect.
We now illustrate this shrinkage-like eect of portfolio weight constraints through an example.
The covariance matrix estimate S is the sample covariance matrix of a random sample of 500 stocks
among all domestic common stocks that have all monthly returns from 1990 to 2000 (132 months).
We consider the eect of imposing the nonnegativity constraint in the global portfolio variance
minimization problem. (The eect of upper bounds on portfolio weights is qualitatively similar
and hence is not reported.) Panel A of Table I reports summary statistics for the row average
covariances of S, the Lagrange multipliers for the nonnegativity constraint, and the row average
covariances of the adjusted estimate

S. The row average covariances of S are the row sums of S
divided by 500. Recall that the eect of the nonnegativity constraint is to reduce each element
in the i-th row of S by
i
. Therefore we report row average covariances instead of row sums of S,
since the former can be compared with the Lagrange multipliers to give a sense of the magnitude
of the adjustment due to the constraint.
6
Insert Table I about here
We can see that the row average covariances of S vary over a wide range: the lowest row average
covariance is 0.27 10
4
, and the highest is 51.86
4
. Thus the range (i.e., the highest minus the
lowest row average covariance) is 52.13 10
4
.
The row average covariances of the adjusted covariance estimate

S are less variable. Their range
is only 44.42 10
4
. So those rows that have the largest covariances tend to be reduced due to the
nonnegativity constraint. This is consistent with our previous analysis.
For comparison purposes, we also reported the row average covariances of the Ledoit estimate
of the covariance matrix S
Ledoit
, which is a shrinkage estimator of the covariance matrix (Ledoit
(1996, 1999)). We can see that the mean, standard deviation, minimum, and maximum of the row
averages of

S are very similar to those of S
Ledoit
. This is consistent with the interpretation that
imposing the nonnegativity constraint has a shrinkage-like eect.
If imposing the nonnegativity constraint is similar to shrinking S toward the one-factor model
as in Ledoits estimator, then we expect changes in row averages from S to

S to be correlated with
changes in row averages from S to S
Ledoit
. Panel B of Table I examines this issue.
We rst regress the changes in row averages from S to

S on the row averages of S, yielding
a highly signicant coecient of 0.143. We then regress the changes in row averages from S to
S
Ledoit
on the row averages of S, obtaining a highly signicant coecient of 0.226. Finally we
regress the changes in row averages from S to

S on the changes in row averages from S to S
Ledoit
,
yielding a highly signicant coecient of 0.296.
The rst regression says that, from S to

S, the highest row averages tend to be reduced the most.
This is consistent with our previous interpretation of the eect of the nonnegativity constraint. The
last regression shows that the eect of imposing the nonnegativity constraint on the row averages
of S is to some extent similar to the eect of shrinking S toward the one-factor model.
Panel C of Table I is similar to Panel B, except that we examine the changes in distinct elements
from S to

S. The results are qualitatively similar to those in Panel B. We emphasize that even
though some of the adjusted R
2
appear low in Panel C, the regressions are highly signicant as
indicated by the F-values and t-statistics.
To summarize, Table I shows that the eect of imposing the nonnegativity constraint is similar
to shrinking S toward the one-factor model. This is consistent with a shrinkage interpretation of
the nonnegativity constraint.
We would like to emphasize that we are not saying that

S is an equally good or better estimator
as S
Ledoit
. Our interest is not to derive a better covariance matrix estimator. Instead, we are trying
to explain why imposing the constraints that are wrong in population helps the out-of-sample
performance of the optimal portfolios in practice. We do that by showing that the constraints have
a shrinkage-like eect on the covariance matrix estimate.
C. The Role of Constraints in Constructing Mean-Variance Ecient Portfolios
Under certain conditions, we can also give a shrinkage-like interpretation to the constraints
7
in constructing mean-variance ecient portfolios. Liu (2001) and an earlier version of this paper
(Jagannathan and Ma (2002)) give more details. The result is summarized as the following. When
S is nonsingular, the eect of the lower bound on portfolio weights is to adjust the mean returns
upward by an amount proportional to the corresponding Lagrange multiplier, and the eect of the
upper bound on portfolio weights is a similar downward adjustment.
From an empirical point of view, this result is less interesting because tangency portfolios
perform worse than the global minimum variance portfolios out of sample, and hence practitioners
are less interested in tangency portfolios when sample mean is used as an estimator of expected
return..
II. A Simulation Study of the Tradeo Between
Specication Error and Sampling Error
In the last section, we argued that imposing the nonnegativity constraint in global portfolio
variance minimization has a shrinkage-like eect. This shrinkage-like eect can reduce sampling
error. On the other hand, the constraint is wrong in population when asset returns have a single
dominant factor, as argued by Green and Hollield (1992). So whether the nonnegativity constraint
can help depends on the tradeo between these two eects: The reduction in sampling error
on the one hand and the increase in the specication error on the other hand. That tradeo
would depend on, among other things, the true covariance structure. Since we do not know the
true covariance structure, we cannot give a quantitative answer that holds true for all underlying
covariance structures in the population. We therefore rely on using Monte Carlo simulation in
this section to further examine the tradeo between the two eects. By comparing the simulation
results with the results using the real data, which are reported in the next section, we are also able
to provide additional insights regarding the true covariance structure of U.S. stock returns.
We assume a two-factor return structure in the simulation.
4
We consider a two-factor structure
because the literature often nds more than one factor in the return structure of U.S. stocks.
Without loss of generality, we set the factor variances to unity and the covariance between them
to zero. We draw the betas for the rst factor for the stocks from an i.i.d normal distribution with
mean one and standard deviation
, which varies from 0 to 0.4 across dierent sets of simulations.

The stocks betas with respect to the second factor are drawn from an i.i.d. normal distribution
with mean zero and standard deviation 0.2. The residual variances are set to be constant over
time, but cross-sectionally, they are drawn from an i.i.d. lognormal (0.8, 0.7) distribution. When
= 0.4, the distribution of the betas with respect to the two factors roughly matches the empirical
distribution of the rst two betas for the NYSE stocks, with the two factors being the rst two of
the ve Connor-Korajczyk factors. Furthermore, in this case, the ratio of residual variance to total
variance matches the same ratio in the NYSE stocks.
5
We allow the number of stocks, N, to range from 30 to 300. For each specication of N and
, we draw the betas and the residual variances for the stocks as specied above. These betas and
residual variances are xed for this specication even as we change the sample size T.
8
Given the sampled betas and residual variances, we then draw returns of dierent sample sizes.
We rst set T = 60 and then vary it from N + 30 to N + 210 by an increment of 60. We allow
T = 60 because this is the sample size used in the simulation with real data in the next section,
and having the same sample size allows us to compare the results.
Given the simulated returns data, we estimate the return covariance matrix, and form the non-
negativity constrained and unconstrained global minimum variance portfolios. We then calculate
the ex post variances of these portfolios using the assumed true covariance matrix. This procedure
is repeated ten times and the averages of ex post variances are reported.
We consider two covariance matrix estimators. The rst is the sample covariance matrix. The
second is the one-factor model. When the covariance matrix is estimated with the one-factor model,
we assume the factor return is known, and is the return of the rst factor used in the simulation.
Before we discuss the eect of imposing the nonnegativity constraint, we refer the reader to
Table II, which gives the short interest in the unconstrained global miminum variance portfolios
constructed from the true covariance matrix. This table gives us a measure about the extent to
which the nonnegativity constraint is violated (i.e., the specication error of the nonnegativity
constraint for each covariance matrix structure.). Two patterns emerge. (1) For a given dispersion
in the betas, as the number of assets increases, the nonnegativity constraint becomes more severely
violated in population. (2) When the number of assets is not too small, the smaller the dispersion
in the betas (while there is some dispersion), the more severely the nonnegativity constraint is
violated. Both patterns can be understood by following Green and Hollields (1992) two-step
procedure in forming the global minimum variance portfolios.
Insert Table II about here
Table III presents a subset of our simulation results.
6
It compares the eects of imposing the
nonnegativity constraint with that of the single-factor constraint. The table reports the percentage
reductions in the ex post standard deviations of the global minimum variance portfolios constructed
by imposing these constraints relative to that of the unconstrained global minimum variance port-
folios constructed from the sample covariance matrix, and relative to the equal-weighted portfolio.
The three panels,
= 0,
= 0.2, and
= 0.4, represent the cases where the nonnegativity

constraint is correct, is severely wrong, and is moderately wrong, respectively. (When there is no
dispersion in the betas, the nonnegativity constraint holds in population.) We report the cases
where the number of stocks is either 30 or 300, and the sample size is either 60 or 360.
Insert Table III about here
Panel A shows that when the nonnegativity constraint is correct, imposing the nonnegativity
constraint always helps. The gain from imposing the nonnegativity constraint is comparable to
that obtained by imposing the one-factor structure. In fact, this is true for all the cases in our
simulation, not only those cases reported in the table.
Across all three panels, the numbers in the last two columns are comparable. This means that
if for some reason the nonnegativity constraint has to be imposed, then whether we use the sample
covariance matrix or the one-factor model to estimate the covariance matrix makes little dierence.
9
Again this is true not only for the cases in Table III but for all the cases in our simulation.
When
> 0, so that the true global minimum variance portfolio has negative weights, the
nonnegativity constraint always hurts out-of-sample performance if we use the one-factor model to
estimate the covariance matrix, and the deteriation in performance is dramatic if the number of
stocks is large. However, imposing the nonnegativity constraint helps when the sample covariance
matrix is used and there are too many stocks relative to the length of the time series of observations
used to estimate the covariance matrix. Again this is true in general and not only for the cases
reported in the table.
The numbers in the last two columns indicate that, paradoxically, the nonnegativity-constrained
portfolios do worse than the equal-weighted index when the nonnegativity constraint is correct
(i.e., Panel A), and do better than the equal-weighted index when the nonnegativity constraint
is wrong (i.e., Panels B and C). This is because when there is dispersion in the betas (so that
the nonnegativity constraint is wrong in population), even if shortsales are not allowed, we can
still reduce the exposure to systematic risk by investing only in the low-beta stocks. Thus we can
achieve lower risk than in the equal-weighted index. On the other hand, when there is no dispersion
in the betas (so that the nonnegativity constraint is correct in population), we can not reduce
systematic risk at all. Hence there is not much dierence between optimal diversication and naive
diversication, even though we know the population covariance matrix. So optimal diversication
based on sample estimates can perform worse than naive diversication due to sampling error.
Based on the results from the full set of simulation, we also nd a meaningful tradeo between
specication error and sampling error when
> 0 and the covariance matrix estimate is the sample

covariance matrix. First, when
= 0.1, the nonnegativity constraint always improves the out-of-

sample performance. On the other hand, as
goes from 0.2 to 0.4, we always nd cuto points for

the sample size T, so that when we have fewer observations than the cuto points, the nonnegativity
constraint improves the out-of-sample performance, and when we have more observations than the
cuto points, the nonnegativity constraint hurts the out-of-sample performance. These cuto points
grow as we increase the number of assets (as expected), but are not sensitive to changes in
. These
cuto points are presented in Table IV. Recall that
= 0.4 roughly corresponds to the covariance

structure of the NYSE stocks, and that
= 0.2 or 0.3 represents cases where the nonnegativity

constraint is more severely violated than when
= 0.4. Since these cuto points are not sensitive

to changes in
in this range, we can treat them as a rough guideline for practitioners who use
U.S. stocks.
Insert Table IV about here
III. The Eect of Portfolio Weight Constraints:
An Empirical Examination
A. Data and Methodology
In this section we examine empirically the eect of portfolio weight constraints. As we have
10
said before, whether these weight constraints help or hurt is an empirical issue and depends on the
specic covariance matrix estimator. For the estimators that have large sampling errors, such as the
sample covariance matrix, the portfolio weight constraints are likely to be helpful, as documented
by Frost and Savarino (1988) and demonstrated again in the simulation study in the last section.
However, for the factor models and shrinkage estimators, the portfolio weight constraints are likely
to be harmful.
We examine the eect of the portfolio weight constraints on the out-of-sample performance of
minimum variance and minimum tracking error portfolios formed using a number of covariance
matrix estimators. This is done following the methodology in Chan, Karceski, and Lakonishok
(1999). At the end of each April from 1968 to 1998, we randomly choose 500 stocks from all
common domestic stocks traded on the NYSE and the AMEX, with stock price greater than ve
dollars, market capitalization more than the 80th percentile of the size distribution of NYSE rms,
and with monthly return data for all the immediately preceding 60 months. We use return data
for the preceding 60 months to estimate the covariance matrix of the returns of the 500 stocks. For
estimators that use daily data, the daily returns during the previous 60 months of the same 500
stocks are used. When a daily return is missing, the equally weighted market return of that day is
used instead.
When variance minimization is the objective, we form three global minimum variance portfolios
using each covariance matrix estimator only two if the covariance matrix estimate is singular. The
rst portfolio is constructed without imposing any restrictions on portfolio weights, the second is
subject to the constraint that portfolio weights should be nonnegative, and the third, in addition,
faces the restriction that no more than two percent (i.e., 10 times of the equal weight) of the
investment can be in any one stock. Each of these portfolios are held for one year. Their monthly
returns are recorded, and at the end of April of the next year, the same process is repeated. This
gives at most three minimum variance portfolios that have post-formation monthly returns from
May 1968 to April 1999 for each covariance matrix estimator. We use the standard deviation of
the monthly returns on these portfolios to compare the dierent covariance matrix estimators.
For tracking error minimization, following Chan, Karceski, and Lakonishok (1999) we assume
the investor is trying to track the return of the S&P 500 index. As in the case of portfolio variance
minimization, we construct three tracking error minimizing portfolios for each covariance estimator.
Notice that constructing the minimum tracking error variance portfolio is the same as constructing
the minimum variance portfolio using returns in excess of the benchmark, subject to the restriction
that the portfolio weights sum to one.
B. Covariance Matrix Estimators
The rst estimator is the sample covariance matrix:
S
N
=
1
T 1
T
t=1
(h
t

h)(h
t

h)
,
where T is the sample size, h
t
is a N 1 vector of stock returns in period t, and

h is the average
11
of these return vectors.
The second estimator assumes that returns are generated according to Sharpes (1963) one-
factor model given by:
r
it
=
i
+
i
r
mt
+
it
,
where r
mt
is the period t return on the value-weighted portfolio of stocks traded on the NYSE,
AMEX, and Nasdaq. Then the covariance estimator is
S
1
= s
2
m
BB
+D. (13)
Here B is the N 1 vector of s, s
2
m
is the sample variance of r
mt
, and D has the sample variances
of the residuals along the diagonal and zeros elsewhere.
.
The third estimator is the optimal shrinkage estimator of Ledoit (1999). It is a weighted average
of the sample covariance matrix and the one-factor model-based estimator:
S
L
=

T
S
1
+ (1

T
)S
N
,
where is a parameter that determines the shrinkage intensity that is estimated from the data.
Ledoit (1999) shows this estimator outperforms the constant correlation model (Elton and Gruber
(1973) and Schwert and Seguin (1990)), the single-factor model, the industry factor model, and
the principal component model with ve principal components.
7
For the fourth set of estimators
we consider the Fama and French (1993) three-factor model, the Connor and Korajczyk (1986,
1988) ve-factor model, and a three-factor version of it which includes only the rst three of the
ve factors. This gives three additional covariance matrix estimators, each corresponding to one of
these multifactor models.
8
Finally we consider several covariance matrix estimators that use daily return data. These in-
clude the daily return sample covariance matrix, daily one-factor model, daily Fama-French three-
factor model, and daily Connor-Korajczyk ve-factor and three-factor models. These models are
similar to the corresponding monthly models. However, we incorporate the corrections for mi-
crostructure eects suggested by Scholes and Williams (1977), Dimson (1979), and Cohen et al.
(1983). We also develop a new estimator for the covariance matrix of returns using daily return
data that nests these three estimators. Details on the estimation of the monthly return covariance
matrix using daily return data that allow for microstructure eects are provided in Appendix B.
C. Empirical Results
Table V gives the characteristics of minimum variance portfolios constructed using various
covariance matrix estimates.
9
Judging by the ex post standard deviations of the optimal portfolios,
the shrinkage estimator proposed by Ledoit (1999) and the sample covariance matrix of daily return
are the best performers. Since the Ledoit estimator is a particular weighted average of the one-
factor model and the sample covariance matrix, we examined whether a simple average of the two
estimators would do equally well. A randomly weighted average of the one-factor model and the
12
sample covariance matrix has an annualized out-of-sample standard deviation of 10.34 percentage
points (not reported in Table V), which is not much dierent from the 10.76 percentage points
for the Ledoit estimator. We thus suspect that the sampling errors associated with the estimated
optimal shrinkage intensity in the Ledoit estimator is rather large. Notice that even for these
two best estimators, the unconstrained global minimum variance portfolios involve taking short
positions that are over 80 percent of the portfolio value.
Insert Table V about here
Next, we turn to the case where no-shortsale restrictions are imposed. Surprisingly, the mini-
mum variance portfolio constructed using the monthly sample covariance matrix compares favorably
with all the other covariance matrix estimators. The out-of-sample annualized standard deviation
is about 12 percentage points per year for all of the estimators, including the shrinkage estimator
of Ledoit and the daily sample covariance matrix. Imposing the no-shortsales restriction leads to a
small increase between eight and 14 percent in the standard deviation of the minimum variance
portfolios constructed using factor models. This decline in the performance is consistent with the
observation by Green and Hollield (1992) that shortsale restrictions probably do not hold in the
population. The number of assets in the portfolio varies from a low of 24 for the monthly sample
covariance matrix estimator to a high of 65 for the daily sample covariance matrix. In compar-
ison, the equally weighted portfolio of the 500 stocks has an annualized standard deviation of 17
percentage points, while the value weighted portfolio of the 500 stocks has a standard deviation
of 16 percentage points. A portfolio of 25 randomly picked stocks has a standard deviation of 18
percentage points, which is 40 percent more than that of the optimal minimum variance portfolio
constructed using the sample covariance matrix subject to the no-shortsale restrictions.
This clearly indicates that portfolio optimization can achieve a lower risk than naive diversi-
cation. Imposing an upper bound of two percentage points on portfolio weights in addition to a
lower bound of zero does not aect the out-of-sample variance of the resulting minimum variance
portfolios in any signicant way.
When daily returns are used, the sample covariance matrix estimator performs the best when
there are no portfolio weight constraints. The corrections for microstructure eects suggested in
the literature do not lead to superior performance. When portfolio weights are constrained to be
nonnegative, all the models perform about equally well.
Table VI presents the performance of the minimum tracking error portfolios. Several patterns
emerge. Among the estimators that use monthly data, the monthly covariance matrix estimator
and the Ledoit estimator perform the best. The former has only about 200 stocks, whereas the
latter has about 300 stocks.
Insert Table VI about here
The tracking error of the one-factor model with nonnegativity constraints is 5.04 percent, which
is rather large compared to 3.36 percent for the sample covariance matrix estimator. This should not
be surprising since the rst dominant factor becomes less important for tracking error minimization.
The portfolio weight constraints are not binding with the one-factor model. This is strong support
13
for the Green and Hollield conjecture that large negative (and positive) weights are due to the
presence of a single dominant factor which is eectively removed when we consider tracking error
minimization.
The multifactor models of Connor and Korajczyk (not reported in the table) as well as Fama
and French perform better than the single-factor model. This is consistent with the observations
in Chan, Karceski, and Lakonishok (1999). As is to be expected, portfolio weight constraints are
not important when using multifactor models. However, the tracking error for the factor models
is about one percentage point more than that for the sample covariance matrix and the Ledoit
estimators.
With daily data we expect the precision of all the estimators to improve. If our conjecture
that portfolio weight constraints lead to better performance due to the shrinkage eect, then such
constraints should become less important when daily data is used. This is what we nd. When
shortsales are allowed, the short interest is 36.6 percent, a substantial amount. Imposing the
nonnegativitiy constraint on portfolio weights reduces the number of stocks to 231. However, the
performance is hardly aected. The tracking error goes down from 2.94 percent to 2.78 percent.
With daily data, the sample covariance matrix estimator dominates all factor models, with the
factor models all perform equally well. Again corrections for microstructure eects make little
dierence in the ex post tracking error performance.
The value-weighted portfolio of the 500 stocks performs the best. This is to be expected since
the benchmark is the S&P500 portfolio, which is value weighted.
Table VII reports the t-tests for the dierence between the mean returns and mean squared
returns of the portfolios, compared with the nonnegativity-constrained portfolio from the monthly
return sample covariance matrix. Since the dierences in mean returns are all insignicant, the
t-test for the dierence in squared returns serves as a test for the dierence in return variances,
which is the focus of our study. We can see that for both portfolio variance minimization and
tracking error minimization, once the nonnegativity constraint is imposed, the more sophisticated
estimators do not in general give better out-of-sample performance than the monthly return sample
covariance matrix. However, there is evidence that using daily returns can lead to smaller out-of-
sample tracking error. Why using daily returns can lead to lower tracking error but not total risk
is an issue for future investigation.
Insert Table VII about here
D. A Comparison of the Global Minimum Variance Portfolios and Tangency Portfolios
Earlier we justied examining the global minimum variance portfolio instead of the tangency
portfolio based on results in the literature. In this section we provide a comparison of the two
portfolios that justies our focus on the global minimum variance portfolios.
Table VIII gives the comparative results for the portfolios constructed using a random sample
of 500 stocks. We only include the result for several covariance matrix estimators that use monthly
return data. (The results for other estimators are qualitatively similar.) The procedure of con-
14
structing the optimal portfolios and examining their out-of-sample characteristics is the same as in
the previous subsections. One dierence is that here we record the ex post excess returns in excess of
the one-month T-bill rate, and compare Sharpe ratios acrosss dierent optimal portfolios. In terms
of out-of-sample Sharpe ratio, all the tangency portfolios perform worse than the equally weighted
portfolio. The unconstrained portfolios perform uniformly worse than the constrained portfolios.
For every covariance matrix estimator, the global minimum variance portfolios perform better than
the tangency portfolios. The in-sample optimism (i.e., the fact that the in-sample Sharpe ratios
are much higher than the out-of-sample ones) of the unconstrained portfolios is rather large, and
the unconstrained tangency portfolios involve substantial short positions that would be dicult
to implement in practice. For example, for the Fama-French three-factor model, the total short
position is on average 704 percent of the portfolios value. These problems are much less severe
for the global minimum variance portfolios.
Insert Table VIII about here
Table IX gives the result for the optimal portfolios constructed using the 25 Fama-French
size/book-to-market sorted portfolios. The procedure of constructing the optimal portfolios is the
same as before, except the primitive assets are the 25 portfolios. Since portfolio average returns
have smaller variances compared to individual stocks, they provide a more accurate estimate of the
expected return on the corresponding portfolios. However, even in this case, for every covariance
matrix estimator the unconstrained global minimum variance portfolios have a higher Sharpe ratio
out of sample than the corresponding unconstrained tangency portfolios. The constrained global
minimum variance portfolios perform worse than the unconstrained ones, whereas the opposite is
true for the tangency portfolios.
Insert Table IX about here
Comparing the results in the two tables, we see that for a large cross-section of stocks, the global
minimum variance portfolio performs better than the tangency portfolio, while for the 25 Fama-
French portfolios, and when the nonnegativity constraint is imposed, the opposite is true. This
justies our focus on the global minimum variance portfolios when working with a large collection
of stocks.
Another issue is that the constrained global minimum variance portfolios, while having a higher
Sharpe ratio than the equally weighted portfolio, are likely to be less robust, in the sense that
something bad could happen to a few assets that aect the global minimum variance portfolios
adversely but leave the equally weighted portfolio almost unaected. This is because the constrained
global minimum variance portfolios have far fewer assets about a sixth as many as the equally
weighted portfolio in the case of the 25 Fama-French portfolios and a twentieth as many assets in
the 500 asset case. This issue needs to be examined in future research.
IV. Concluding Remarks
Practitioners often restrict portfolio weights to be positive when constructing minimum variance
15
and minimum tracking error portfolios. Based on the covariance structure of U.S. stock returns
Green and Hollield (1992) argue that minimum variance portfolios will contain extreme positive
and negative weights. In that case restricting portfolio weights to be positive as done in prac-
tice should hurt whereas the empirical evidence is to the contrary. We reconcile this apparent
contradiction.
We show that constructing a minimum risk portfolio subject to the constraint that portfolio
weights should be positive is equivalent to constructing it without any constraints on portfolio
weights after modifying the covariance matrix in a particular way. The modication typically
shrinks the larger elements of the covariance matrix towards zero. This has two eects. On the
one hand, to the extent an estimated large covariance is due to sampling error the shrinkage leads
to a more precise estimate of the corresponding element in the population. On the other hand, the
population covariance itself could be large in which case the shrinkage introduces specication error.
The net eect depends on the tradeo between sampling and specication errors. If the sampling
error is relatively large when compared to the specication error restricting portfolio weights to be
positive can help even when the restrictions do not hold in the population. Imposing upper bounds
on portfolio weights while constructing minimum risk portfolios have similar shrinkage-like eects.
We examine the tradeo between sampling error and specication error using simulations.
We demonstrate through Monte Carlo simulations calibrated to U.S. stock return characteristics
that the benet to imposing non-negativity constraints on portfolio weights can be substantial in
large cross sections even when the constraints are wrong. When the no short-sale restriction is
already in place, minimum variance and minimum tracking error portfolios constructed using the
sample covariance matrix perform as well as those constructed using covariance matrices estimated
using factor models and shrinkage methods. The use of daily return data instead of monthly return
data helps when constructing minimum tracking error portfolios but not in constructing minimum
variance portfolios. Corrections for market microstructure eects that have been suggested in the
literature when using daily return data do not help.
A striking feature of minimum variance portfolios constructed subject to the restriction that
portfolio weights should be nonnegative is that investment is spread over only a few stocks. The
minimum variance portfolio of a 500 stock universe has between 24 and 40 stocks depending on
which covariance matrix estimator was used. The annualized standard deviation of the return on
the minimum variance portfolio constructed using the sample covariance matrix and 60 months of
observations on returns is 12.43 percent. In contrast the return on the equally weighted portfolio
of all the 500 stocks in the universe has a standard deviation of 17.48 percent, i.e., 1.4 times as
large. Also, the minimum variance portfolio also has a higher sample Sharpe ratio than the equally
weighted portfolio of the 500 stocks in the universe. The return on a portfolio of 25 randomly picked
stocks has a standard deviation of 17.78 percent, about the same as the equally weighted portfolio
of 500 stocks in the universe. While there is little incremental benet to naive diversication
beyond 25 stocks, there is substantial benet to picking the right 25 stocks using standard portfolio
optimization methods.
Having only 25 stocks in the minimum variance portfolio should raise some concern. Variance
16
is an adequate measure of risk of a return that has a Normal distribution. The return on a portfolio
of a large collection of stocks will in general be close enough to Normal in distribution to justify the
use of variance as the measure of risk. This may not be true for a 25 stock portfolio. In particular
the probability of an extreme event, both good as well bad, may be substantially higher than that
computed by assuming a Normal distribution. Imposing upper bounds on portfolio weights would
be one way to ensure that optimal portfolios will contain a suciently large collection of stocks.
17
Appendix A: Proofs
Proof of Proposition 1:
10
The matrix

S is obviously symmetric. Now we prove that it is positive semi-denite. Suppose
that
(
1
, ,
N
,
1
, ,
N
,
1
, ,
N
,
0
) (
,
0
)
is a solution to the constrained portfolio variance minimization problem (1) - (4). For any vector
x,
x

Sx = x
Sx x
(1
+1
)x +x
(1
+1
)x = x
Sx 2(x
1)(x
( )). (A1)
(Here 1 is a vector of ones.) By the rst order condition, = S
0
1. Hence, x
( ) =
x
S
0
x
1. Therefore
2(x
1)(x
( )) = 2(x
1)(x
S) 2
0
(x
1)
2
. (A2)
But
|(x
1)(x
S)| = |(x
1)(x
S
1/2
)(S
1/2
)| |(x
1)|(x
Sx)
1/2
(
S)
1/2
.
The equality holds since S is positive semi-denite, and the weak inequality is due to the Cauchy-
Schwarz inequality.
Again from the rst order condition,
0
S =
+
0
1 =
0

1
0
.
So
|(x
1)(x
S)| |(x
1)|(x
Sx)
1/2
(
0
)
1/2
.
Combining the above inequality with (A1) and (A2), we have
x

Sx = x
Sx 2(x
1)x
S + 2
0
(x
1)
2
x
Sx 2|(x
1)x
S| + 2
0
(x
1)
2
x
Sx 2|(x
1)|(x
Sx)
1/2
(
0
)
1/2
+ 2
0
(x
1)
2
= (a b)
2
+b
2
, (A3)
where a = (x
Sx)
1/2
and b =
1/2
0
|(x
1)|. Obviously this is always nonnegative. So

S is positive
semi-denite.
Because

S is positive semi-denite, to show that = (
1
, ,
N
) is a (unconstrained) global
minimum variance portfolio of

S, it suces to verify the rst order condition:
S = S (1
+1
) + (1
+1
)
= S 1
+1 (
1) +1
= S + (
1)1 +
= (
0
+
1)1.
18
The second equality follows from the fact that
i
i
= 0 for all i, and
i
(
i
) = 0 for all i.
The third equality holds because

j
w
j
= 1, and the last equality follows from (5). The fact that
S = (
0
+
1)1 shows that solves the (unconstrained) variance minimization problem for
covariance matrix

S.
QED
Proof of Proposition 2:
First, lets re-write the second constraint in the constrained MLE problem as
(1 )
i,j

k=i
k,j
.
Let the constrained MLE be

. The Kuhn-Tucker conditions for the constrained MLE problem are
(Morrison (1990)):
l
i,i
=
1
2

i,i

1
2
S
i,i
=
i
+ (1 )
i
, all i (A4)
l
i,j
=

i,j
S
i,j
= (
i
+
j
) + (1 )(
i
+
j
), all i < j (A5)
i
0, and
i
= 0 if
i,j
> 0, (A6)
i
0, and
i
= 0 if (1 )
i,j
<
k=i
k,j
. (A7)
These conditions imply that the constrained MLE can be written as
= S (1
+1
) + (1 )(1
+1
)
Notice that

has the same form as

S.
We will only prove Part 1. The proof of Part 2 is similar.
Let {
i
,
i
,
i
}
N
i=1
be a solution to the constrained portfolio variance minimization problem (1)
- (4), given S, and construct

S according to (8). Then we can easily verify that {
i
, (1 )
i
}
N
i=1
and

S satises (A4), (A5), and the rst halves of (A6) and (A7). Now we need to verify the second
halves of (A6) and (A7), i.e.,

j
S
i,j
> 0 implies
i
= 0 and (1 )
S
i,j
<
k=i
S
k,j
implies
i
= 0. By Proposition 1, is the unconstrained global minimum variance portfolio of

S,
so
i
=

j
S
i,j
/
S
j,k
. From the second half of (6) we know that
i
> 0 implies
i
= 0. This
says that

j
S
i,j
> 0 implies
i
= 0. Likewise, the second half of (7) implies the second half of
(A7). QED
Appendix B: Covariance Matrix Estimators That Use
Daily Returns
This appendix describes how we estimate monthly covariance matrices using daily return data
after taking into account serial correlations and cross-correlations at various leads and lags induced
by microstructure eects. We review the estimators proposed by Scholes and Williams (1977),
19
Dimson (1979), and Cohen et al. (1983, hereafter CHMSW). We point out that the CHMSW
estimator of the sample covariance matrix is usually not positive semi-denite. We then propose a
new methodology to estimate the covariance matrices of monthly returns using daily returns that
are always positive semi-denite. Our estimator uses the fact that the continuously componded
monthly return is the sum of continuously compounded daily returns.
The CHMSW estimator is based on the following relationship:
cov(r
t
j,t
, r
t
k,t
) = cov(r
j,t
, r
k,t
) +
L
n=1
cov(r
j,t
, r
k,tn
)
+
L
n=1
cov(r
j,tn
, r
k,t
), (A8)
for any pair of stocks j and k, j = k. Here r
t
denotes the true date t return and r denotes the
observed return; t and t n denote dates t and t n. Based on this relation we can estimate
cov(r
t
j,t
, r
t
k,t
) using observed daily returns. In practice, L is set to three and the variances are esti-
mated using sample variances without any adjustment (Cohen et al. (1983) and Shanken (1987)).
The Scholes-Williams estimator is the special case of the above, with L set to one. The estimate
of the full sample covariance matrix using the CHMSW method is denoted as Sample Covariance
Matrix (CHMSW) in Tables V-VII.
Equation (A8) is also valid if either asset (or both) is a portfolio. For a well-diversied portfolio,
(A8) is approximately valid for its variance also. Based on this, we can estimate a stocks beta as
its covariance with the market portfolio divided by the market portfolios variance. We estimated
daily one-factor models using this strategy but the results for these estimators are not reported for
brevity.
There is a problem with these estimators. The estimated covariance matrix, constructed from
individual covariances and variances, is not positive semi-denite. This is problematic for portfolio
optimization. We propose a new estimator that does not have this problem.
11
Notice that monthly
log returns are simply sums of daily log returns:
r
i,
=
m
t=(1)m+1
r
i,t
.
Here is month , t is day t, and m is the number of days in a month. Then the monthly return
covariance is
cov(r
i,
, r
j,
) =
m
t=(1)m+1
m
s=(1)m+1
cov(r
i,t
, r
j,s
).
Unlike the CHMSW estimator (A8), the above is valid for any i and j, even if either one (or both)
20
is a portfolio. Assuming covariance stationarity as usual, we can drop the time subscripts, and get
cov
M
(r
i
, r
j
) = m cov
D
(r
i,t
, r
j,t
)
+ (m1) (cov
D
(r
i
, r
j,t+1
) + cov
D
(r
i,t+1
, r
j,t
))
+ (m2) (cov
D
(r
i
, r
j,t+2
) + cov
D
(r
i,t+2
, r
j,t
))
+
+ (cov
D
(r
i,t
, r
j,t+m1
) + cov
D
(r
i,t+m1
, r
j,t
)). (A9)
The superscripts M and D denote the covariance of the monthly returns and the covariance of the
daily returns, respectively. We set m = 21 for there are about 21 trading days in one month.
An obvious approach is to use the sample counterpart of the RHS of (A9) to estimate the
monthly return covariance. However, to guarantee that the covariance matrix is positive semi-
denite, we need to further adjust the covariance estimates (A9) slightly. Let
R
0
= (r
t,i
r
.,i
)
t=1,...,T; i=1,...,N
,
represent the matrix of demeaned returns. For j = 1, , m 1, let R
j
be the same size as R
0
,
with the rst j rows set to zeros, and the remaining T j rows the same as R
0
s rst T j rows
(i.e., R
j
is the matrix of lag-j demeaned returns). Let
j
= R
0
R
j
/T,
S = m
0
+
m1
j=1
(mj)(
j
+
j
). (A10)
Then S is positive semi-denite (see Newy and West (1987) for the proof) and S is a consistent
estimator of the RHS of (A9). The covariance matrix estimated this way is denoted as Sample
Covariance Matrix (New) in Tables V - VII.
If we assume a kfactor model, then the beta estimates are
b = ( var(f))
1
cov(f, r).
Here b is k N, var(f) is the estimated factor covariance matrix (of size k k) and is estimated
according to (A10), and cov(f, r) is estimated (similar to (A10)) by
cov(f, r) =
1
T
mF
0
R
0
+
m1
j=1
(mj)(F
0
R
j
+F
j
R
0
)
,
with F
j
dened similarly as R
j
.
For factor models, the residual covariance matrix is assumed to be diagonal and the residual
variances are estimated by the sample variances of residuals calculated from observed stock returns,
observed factor returns, and the estimated betas. In Tables V through VII, the estimators denoted
as Daily Fama-French 3-Factor Model are estimated using this strategy. We also examined
21
the performance of a daily one-factor model, a daily Connor-Korajczyk three-factor model, and a
daily Connor-Korajczyk ve-factor model, all estimated using this strategy. These results are not
reported for brevity.
We construct the daily Connor-Korajczyk factors and Fama-French factors by following the
same procedure as outlined in Connor and Korajczyk (1988) (using daily returns instead of monthly
returns) and Fama and French (1993).
22
REFERENCES
Best, Michael J., and Robert R. Grauer, 1991, On the sensitivity of mean-variance-ecient portfo-
lios to changes in asset means: Some analytical and computational results, Review of Financial
Studies 4, 315-342.
Black, Fisher, and Robert Litterman, 1992, Global portfolio optimization, Financial Analysts Jour-
nal, 28-43.
Bloomeld, Ted, Richard Leftwich, and John Long, 1977, Portfolio strategies and performance,
Journal of Financial Economics 5, 201-218.
Chan, Louis K. C., Jason Karceski, and Josef Lakonishok, 1999, On portfolio optimization: Fore-
casting covariances and choosing the risk model, Review of Financial Studies 12, 937-974.
Cohen, Kalman, Gabriel Hawawini, Steven Maier, Robert Schwartz, and David Whitcome, 1983,
Friction in the trading process and the estimation of systematic risk, Journal of Financial
Economics 12, 263-278.
Connor, Gregory, and Robert A. Korajczyk, 1986, Performance measurement with the arbitrage
pricing theory, Journal of Financial Economics, 15, 373-394.
Connor, Gregory, and Robert A. Korajczyk, 1988, Risk and return in an equilibrium APT: Appli-
cations of a new test methodology, Journal of Financial Economics 21, 255-289.
Dimson, Elroy, 1979, Risk measurement when shares are subject to infrequent trading, Journal of
Financial Economics 7, 197-226.
Fama, Eugene, and Kenneth French, 1993, Common risk factors in the returns on stocks and bonds,
Journal of Financial Economics 33, 3-56.
Frost, Peter A., and James E. Savarino, 1986, An empirical Bayes approach to ecient portfolio
selection, Journal of Financial and Quantitative Analysis 21, 293-305.
Frost, Peter A., and James E. Savarino, 1988, For better performance: Constrain portfolio weights,
Journal of Portfolio Management 15, 29-34.
Green, Richard C., and Burton Hollield, 1992, When will mean-variance ecient portfolios be
well diversied? Journal of Finance 47, 1785-1809.
Jagannathan, Ravi, and Tongshu Ma, 2002, Risk reduction in large portfolios: Why imposing the
wrong constraints helps, NBER working paper 8922.
Jobson, J. D., and Bob Korkie, 1980, Estimation for Markowitz ecient portfolios, Journal of the
American Statistical Association 75, 544-554.
Jobson, J. D., and Bob Korkie, 1981, Putting Markowitz theory to work, Journal of Portfolio
Management, 70-74.
23
Jorion, Philippe, 1985, International portfolio diversication with estimation risk, Journal of Busi-
ness 58, 259-278.
Jorion, Philippe, 1986, Bayes-Stein estimation for portfolio analysis, Journal of Financial and
Quantitative Analysis 21, 279-292.
Jorion, Philippe, 1991, Bayesian and CAPM estimators of the means: Implications for portfolio
selection, Journal of Banking and Finance 15, 717-727.
Ledoit, Olivier, 1996, A well-conditioned estimator for large dimensional covariance matrices, Work-
ing paper, UCLA.
Ledoit, Olivier, 1999, Improved estimation of the covariance matrix of stock returns with an appli-
cation to portfolio selection, Working paper, UCLA.
Liu, Ludan, 2001, Portfolio constraints and shrinkage estimation, Working paper, Carroll School of
Management, Boston College.
Markowitz, H. M., 1952, Portfolio selection, Journal of Finance 7, 77-91.
Markowitz, H. M., 1959, Portfolio Selection: Ecient Diversication of Investments (Yale Univer-
sity Press, New Haven, CT).
Michaud, Richard O. 1989, The Markowitz optimization enigma: Is optimized optimal? Financial
Analysts Journal, 31-42.
Morrison, Donald F., 1990, Multivariate Statistical Methods (McGraw-Hill, New York).
Newy, Whitney K, and Kenneth D. West, 1987, A simple, positive semi-denite, heteroskedasticity
and autocorrelation consistent covariance matrix, Econometrica 55, 703-708.
Schwert, G. Williams, and Paul L. Seguin, 1990, Heteroskedasticity in Stock Returns, Journal of
Finance 45, 1129-1155.
Scholes, Myron, and Joseph Williams, 1977, Estimating betas from nonsynchronous data, Journal
of Financial Economics 5, 309-328.
Shanken, Jay, 1987, Nonsynchronous data and the covariance-factor structure of returns, Journal
of Finance 42, 221-231.
Sharpe, William, 1963, A simplied model for portfolio analysis, Management Science 9, 277-293.
24
Footnotes
1
See Frost and Savarino (1988) for an excellent discussion.
2
Using monthly stock index return data for G7 countries Jorion (1985) convincingly argues that
benets from diversication are more likely to accrue from a reduction in risk. In the data set
he examined, the global minimum variance portfolio had the best out-of-sample performance.
It outperformed classical tangent portfolio, the tangent portfolio constructed using the Bayes-
Stein estimator for the vector of mean returns, and the value weighted and equally weighted
portfolios. Using simulation methods, Jorion (1986) showed that this conclusion is robust
if the sample size is not large. Jorion (1991) found that the minimum variance portfolio
constructed using returns on seven industry stock index portfolios performed as well as the
CRSP equally weighted and value weighted stock indices during the January 1926 to December
1987 period, in out-of-sample tests. The performance was comparable to that of the tangent
portfolio constructed using the Bayes-Stein estimator for the mean. Bloomeld, Leftwich,
and Long (1977) found that portfolios constructed using mean-variance optimization did not
dominate an equally weighted portfolio. On the other hand, Chan, Karceski, and Lakonishok
(1999) documented that the constrained global minimum variance portfolios outperform the
equally weighted portfolio.
3
It is well recognized in the literature that imposing portfolio weight constraints leads to supe-
rior out-of-sample performance of mean-variance ecient portfolios. However, to our knowl-
edge no one has noticed that the performance improvement is so large that it is comparable
to that attained using the other alternatives.
4
We also performed a simulation with a one-factor structure. The results are qualitatively
similar. These results are available upon request.
5
When we scale the factor variances and residual variances up or down by a common factor, the
population and sample covariance matrices are also scaled up or down by the same factor, and
hence optimal portfolio weights will stay the same, as will the relative performance of dierent
portfolios. Therefore, the discussion below is not aected by the fact that we normalized the
factor variances to one.
6
To save space, we do not report the full result of the simulations, which is available upon
request.
7
For tracking error variance minimization, we also considered Ledoits (1996) estimator that
shrinks the sample covariance matrix toward the identity matrix. The results are similar and
we do not report them.
8
For tracking error variance minimization, the loadings on the rst factor in these multifactor
models are set to zero for every stock.
25
9
To save space, we only report a subset of the simulation results in Tables V to VII. The results
of using Connor-Korajczyk three- and ve-factor models are very similar to those using the
Fama-French three-factor model. This is true when daily returns are used as well. So these
results are not reported. Several versions of the daily one-factor model are dominated by the
daily return sample covariance matrix. So these results are not reported either.
10
We thank Gopal Basak for providing a part of this proof.
11
There is a dierence between our approach and the approach taken by Scholes and Williams
(1977) and CHMSW. While they want to estimate the true covariances and betas using the
daily returns, we want to use daily returns to estimate the covariances and betas of monthly
returns.
26
Table I
The Eect of Portfolio Weight Constraints
This table shows the eects of nonnegativity constraints on portfolio weights. The covariance matrix
S is the sample covariance matrix of monthly returns of a random sample of 500 stocks from 1990
to 2000 (132 months). The variable is the vector of Lagrange multipliers on the nonnegativity
constraint. The matrix

S is our shrunk version of S, S
Ledoit
is the Ledoit estimate of the covariance
matrix. The numbers below the regression coecients are t-statistics.
Panel A: Summary statistics of the row average covariances of S, unconstrained portfolio
weights, and the Lagrange multipliers (all numbers are scaled by a factor of 10
4
.)
Variable Mean Std Dev Minimum Maximum Max Min
Row average covariances 17.95 9.39 -0.27 51.86 52.13
of S
i
s 2.31 1.87 0 7.93 7.93
of

S
of S
Ledoit
Panel B: Regressions of changes in row averages from S to

S and from S to S
Ledoit
(the intercept terms are omitted for brevity)
Independent Variable Row averages Row averages Adj. R
2
F-value
of S of (S
Ledoit
S) (Signicance)
Row averages of (
S S) -0.143 0.514 527.7

(-22.97) (0.000)
Row averages of (S
Ledoit
S) -0.226 0.469 442.5
(-21.04) (0.000)
Row averages of (
S S) 0.296 0.239 157.6

(12.55) (0.000)
Panel C: Regressions of the changes in the distinct elements from S to

S and from S to S
Ledoit
(the intercept terms are omitted for brevity)
Independent Variable S
ij,ji
(S
Ledoit
S)
ij,ji
Adj. R
2
F-value
(Signicance)
(
S S)
ij,ji
-0.0271 0.099 13810
(-117.5) (0.000)
(S
Ledoit
S)
ij,ji
-0.336 0.536 144599
(-380.3) (0.000)
(
S S)
ij,ji
0.0283(54.2) 0.023 2933.4
(54.2) (0.000)
27
Table II
Short Interests in the Global Minimum Variance Portfolios
The true return covariance has a two-factor structure. Across stocks, the betas with respect to the
rst factor is normally distributed with mean one and standard deviation
, the betas with respect

to the second factor are normally distributed with mean zero and standard deviation 0.2. The two
factor returns have unit variances and zero covariance. The stocks residual variances are constant
over time, but follow a lognormal(0.8, 0.7) distribution cross-sectionally. The number of stocks is
N. For each (
, N) pair, we draw the betas and the residual variances according to the above-
mentioned distributional assumptions, and then calculate the true stock return covariance matrix.
We then calculate the unconstrained global minimum variance portfolio and its short interest (i.e.,
the total short positions). This procedure is repeated ten times and the averages are reported
below. The short interests give the extent to which the nonnegativity constraint is wrong.
N
0.1 0.2 0.3 0.4

30 -0.181 -0.401 -0.391 -0.350
60 -0.554 -0.706 -0.584 -0.468
120 -1.188 -1.012 -0.737 -0.515
180 -1.578 -1.161 -0.828 -0.555
240 -1.845 -1.250 -0.816 -0.534
300 -2.042 -1.308 -0.831 -0.526
28
Table III
Eects of Imposing Nonnegativity Constraints and the Single-factor Constraint
This table shows the percentage reduction in the ex post standard deviations of the global minimum
variance portfolios when the nonnegativity and/or single factor model constraints are imposed,
relative to the ex post variances of the global minimum variance portfolios constructed from the
sample covariance matrix and the equal-weighted portfolio. The letter N is the number of assets; T is
the sample size used to estimate the covariance matrix. The value
is the cross-sectional standard

deviation in the stocks betas with respect to the rst factor. When
is 0, the nonnegativity
constraint is correct in population; when
is 0.4, the nonnegativity constraint is violated roughly

to the same extent as using the NYSE stocks; and when
is 0.2, the nonnegativity constraint is

more severely violated than when
is 0.4. The rows with T = Innity are the population results.

N T Percentage reduction in ex post standard deviation of the optimal portfolios
from imposing the constraints
Relative to the portfolio constructed Relative to the equal-weighted
from the sample covariance matrix portfolio
Single- Nonnegativity Both Single- constraint Both
factor constraint factor constraint
model
Panel A:
= 0
30 60 21.6 25.2 26.0 -12.2 -7.0 -6.0
30 360 0.8 1.2 1.4 -1.7 -1.3 -1.0
30 Innity 0.0 0.0 0.0 1.1 1.1 1.1
300 60 na na na -8.1 -11.6 -6.9
300 360 54.6 58.8 59.1 -14.6 -3.9 -3.2
300 Innity 0.0 0.0 0.0 0.2 0.2 0.2
Panel B:
= 0.2
30 60 20.7 16.2 18.7 8.4 3.3 6.1
30 360 1.7 -3.2 -2.6 14.8 10.6 11.1
30 Innity -0.3 -5.6 -5.8 16.8 12.4 12.3
300 60 na na na 45.9 18.1 23.2
300 360 56.2 26.1 26.5 60.3 33.0 33.3
300 Innity -5.0 -80.7 -81.3 62.8 35.9 35.7
Panel C:
= 0.4
30 60 27.2 16.8 21.2 38.9 30.2 33.9
30 360 2.3 -7.6 -7.1 42.6 36.8 37.0
30 Innity -0.1 -10.0 -10.0 43.3 37.7 37.7
300 60 na na na 72.9 50.8 57.6
300 360 59.2 25.0 27.2 79.3 61.9 63.0
300 Innity -0.8 -79.4 -80.7 79.9 64.2 64.0
29
Table IV
Cuto Points for the Sample Sizes When the Nonnegativity Constraints
Start to Hurt the Out-of-Sample Performance
The true return covariance has a two-factor structure. Across stocks, the betas with respect to the
rst factor are normally distributed with mean one and standard deviation db. The number of stocks
is N, and the sample size is T. We estimate the covariance matrix for the stocks using simulated
returns of sample size T using the sample covariance matrix. Then we form the nonnegativity
constrained and unconstrained global minimum variance portfolios using the estimated covariance
matrix, and calculate the ex post variances of these portfolios. For each (
, N) pair, we report the

cuto point for T, such that when the sample size is greater than or equal to the cuto point, the
nonnegativity constraint starts to hurt the out-of-sample performance. An entry of na means such
cuto point is not found in that set of simulation.
N
0.2 0.3 0.4

30 na 120 120
60 150 150 150
120 270 270 330
180 330 270 330
240 390 390 450
300 450 450 450
30
Table V
Ex Post Mean, Standard Deviation, and Other Characteristics
of the Global Minimum Variance Portfolios
At the end of April each year from 1968 to 1997, the covariance matrix of a random sample of 500
stocks is estimated according to various estimators. We use these covariance matrix estimates to
construct the global minimum variance portfolios, both constrained and unconstrained. We hold
the portfolios for the next 12 months and their monthly returns are recorded. The ex post means,
standard deviations, and other characteristics of these portfolios are reported. The C after an
estimator indicates the nonnegativity constrained portfolios, and D after an estimator indicates
the portfolio with both the nonnegativity constraint and an upper bound of two percent. For
the equal-weighted and value-weighted portfolios of 25 stocks, the 25 stocks are randomly selected
from the 500 ones. Means and standard deviations are in percentage per year; maximum weight,
minimum weight, and short interest are in percentage.
Covariance matrix estimator Mean Std Max. Min. Short No. of
Dev weight weight interest positive
weights
Panel A: Covariance matrix estimated using monthly return data
Sample covariance matrix, C 13.55 12.43 18.4 0 0 24.1
Sample covariance matrix, D 13.55 12.85 2.0 0 0 59.5
One-factor model 13.99 11.69 3.8 -0.9 -50.7 268.8
One-factor model, C 12.68 12.62 11.3 0 0 39.3
One-factor model, D 13.51 12.50 2.0 0 0 63.1
Ledoit 13.09 10.76 4.9 -1.7 -80.6 283.2
Ledoit, C 12.79 12.29 13.4 0 0 39.7
Ledoit, D 13.49 12.43 2.0 0 0 65.4
Fama-French 3-factor model 13.04 11.35 4.2 -1.4 -63.5 284.1
Fama-French 3-factor model, C 12.65 12.38 12.2 0 0 40.1
Fama-French 3-factor model, D 13.34 12.53 2.0 0 0 64.0
31
weights
Panel B: Covariance matrix estimated using daily return data
Sample covariance matrix 14.06 10.64 6.3 -2.5 -122.4 270.7
Sample covariance matrix, D 14.22 12.28 2.0 0 0 81.1
Sample covariance matrix of CHMSW, C 13.89 12.31 9.1 0 0 62.5
Sample covariance matrix of CHMSW, D 14.18 12.25 2.0 0 0 80.1
Sample covariance matrix (New) 13.94 10.60 6.3 -2.6 -128.5 269.5
Sample covariance matrix (New), C 13.81 12.26 9.0 0 0 62.9
Sample covariance matrix (New), D 14.12 12.21 2.0 0 0 79.5
Ledoit 14.30 10.52 4.5 -1.1 -57.4 275.6
Ledoit, C 13.79 12.30 7.7 0 0 71.0
Ledoit, D 14.18 12.21 2.0 0 0 84.5
Fama-French 3-factor 13.07 11.25 4 -1.4 -66.9 276.8
Fama-French 3-factor, C 13.03 12.03 12.7 0 0 40.7
Fama-French 3-factor, D 13.09 12.15 2 0 0 63.6
Panel C: Naive diversication
Equal-weighted portfolio of the 500 stocks 14.52 17.48 0.2 0 0 500
Value-weighted portfolio of the 500 stocks 13.39 15.60 7.3 0 0 500
Equal-weighted portfolio of 25 stocks 15.16 17.78 4 0 0 25
Value-weighted portfolio of 25 stocks 14.25 17.79 30 0 0 25
32
Table VI
Ex Post Mean, Standard Deviation, and other Characteristics
of the Minimum Tracking Error Variance Portfolios
At the end of April each year from 1968 to 1997, the covariance matrix of a random sample of
500 stocks is estimated according to various estimators. We use these covariance matrix estimates
to construct the minimum tracking error variance portfolios, both constrained and unconstrained.
The target is the S&P 500 returns. We hold the portfolios for the next 12 months and their monthly
tracking errors are recorded. The ex post means and standard deviations of the tracking errors and
some characteristics of the tracking portfolios are reported. The C after an estimator indicates the
nonnegativity constrained portfolios, and D after an estimator indicates the portfolio with both the
nonnegativity constraint and an upper bound of two percent. For the equal-weighted and value-
weighted portfolios of 25 stocks, the 25 stocks are randomly selected from the 500 ones. Means
and standard deviations are in percentage per year; maximum weight, minimum weight, and short
interest are in percentage.
weights
Panel A: Covariance matrix estimated using monthly return data
Sample covariance matrix, C 4.28 3.36 2.4 0 0 200
Sample covariance matrix, D 4.23 3.40 1.8 0 0 194
One-factor model 4.91 5.04 0.9 0 0 499
One-factor model, C 4.91 5.04 0.9 0 0 499
One-factor model, D 4.91 5.04 0.9 0 0 499
Ledoit 4.18 3.48 1.8 -0.3 -8.7 391
Ledoit, C 4.21 3.34 2.2 0 0 314
Ledoit, D 4.27 3.36 1.8 0 0 312
Fama-French 3-factor model 3.76 4.41 1.4 -0.2 -3.3 433.1
Fama-French 3-factor model, C 3.71 4.39 1.5 0 0 394.6
Fama-French 3-factor model, D 3.71 4.39 1.5 0 0 394.5
33
weights
Panel B: Covariance matrix estimated using daily return data
Sample covariance matrix 3.69 2.94 5.3 -1 -36.6 322.4
Sample covariance matrix, D 4.17 2.96 2 0 0 227.6
Sample covariance matrix of CHMSW, C 3.92 2.75 5.4 0 0 231.4
Sample covariance matrix of CHMSW, D 4.17 2.92 2 0 0 226.3
Sample covariance matrix (New) 3.72 2.92 5.3 -1 -36.8 322.2
Sample covariance matrix (New), C 3.93 2.73 5.4 0 0 228
Sample covariance matrix (New), D 4.17 2.89 2 0 0 224.2
Ledoit 4.20 3.31 3.7 -0.4 -11.1 375.0
Ledoit, C 4.31 3.30 3.8 0 0 317.4
Ledoit, D 4.39 3.41 2.0 0 0 311.1
Fama-French 3-factor 3.82 4.50 1.3 -0.2 -3.7 427.8
Fama-French 3-factor, C 3.76 4.50 1.5 0 0 384.1
Fama-French 3-factor, D 3.77 4.51 1.4 0 0 383.8
Panel C: Naive diversication
Equal-weighted portfolio of the 500 stocks 4.92 6.58 0.2 0 0 500
Value-weighted portfolio of the 500 stocks 3.6 2.37 6.7 0 0 500
Equal-weighted portfolio of 25 stocks 4.66 8.62 4 0 0 25
Value-weighted portfolio of 25 stocks 4.13 8.24 29.7 0 0 25
34
Table VII
T-tests of Equal Mean Returns and Equal Mean Squared Returns
This table reports the t-tests of equal mean returns and equal mean squared returns of the minimum
variance and minimum tracking error variance portfolios. For each such portfolio, we test whether
its mean returns and mean squared returns are statistically dierent from those of the nonnegativity
constrained portfolio constructed from the sample covariance matrix of monthly returns.
Covariance matrix estimator Minimum variance Minimum tracking error
portio portfolio
Equality Equality in Equality Equality in
in mean mean squared in mean mean squared
return return return return
Monthly sample covariance matrix, D -0.01 1.36 -0.53 1.49
One-factor model 0.26 -1.16 1.06 5.35
One-factor model, C -0.76 0.19 1.06 5.34
One-factor model, D -0.05 0.15 1.06 5.34
Ledoit -0.40 -2.99 -0.32 1.06
Ledoit, C -1.05 -0.60 -0.30 -0.32
Ledoit, D -0.07 -0.02 -0.03 0.08
Fama-French 3-factor model -0.38 -2.34 -1.02 4.46
Fama-French 3-factor model, C -1.08 -0.29 -1.15 4.42
Fama-French 3-factor model, D -0.22 0.23 -1.15 4.41
Daily sample covariance matrix 0.38 -3.86 -1.30 -3.29
Daily sample covariance matrix, C 0.35 -0.11 -0.87 -4.43
Daily sample covariance matrix, D 0.65 -0.22 -0.28 -3.26
Daily sample cov matrix of CHMSW, C 0.29 -0.20 -0.89 -4.82
Daily sample cov matrix of CHMSW, D 0.62 -0.31 -0.29 -3.72
Daily sample cov matrix (New) 0.29 -3.82 -1.24 -3.40
Daily sample cov matrix (New), C 0.23 -0.35 -0.87 -5.09
Daily sample cov matrix (New), D 0.57 -0.44 -0.29 -4.11
Daily Ledoit 0.62 -4.34 0.04 -1.10
Daily Ledoit C 0.22 -0.23 0.30 -1.04
Daily Ledoit D 0.60 -0.41 0.50 -0.33
Daily Fama-French 3-factor -0.31 -2.41 -0.82 4.28
Daily Fama-French 3-factor, C -0.51 -0.95 -0.95 4.25
Daily Fama-French 3-factor, D -0.44 -0.78 -0.93 4.26
Equal-weighted portfolio of the 500 stocks 0.54 6.79 0.73 6.76
Value-weighted portfolio of the 500 stocks -0.10 5.25 -1.48 -7.33
Equal-weighted portfolio of 25 stocks 0.82 6.94 0.29 9.48
Value-weighted portfolio of 25 stocks 0.32 7.07 -0.10 9.88
35
Table VIII
The Ex Post Performance of the Tangency and Global Minimum Variance Portfolios
At the end of April each year from 1968 to 1997, the monthly return covariance matrix of a random sample of 500 stocks is estimated
according to various estimators and using returns of the past ve years. The tangency and the global minimum variance portfolios are
then formed and held for one year. The table reports characteristics of in- and out-of-sample excess returns of these portfolios. Means
and standard deviations are those of annual excess returns in excess of the one-month T-bill rate. Average total short position is in
percentages. The term constrained after a covariance matrix estimator means the nonnegativity constrained optimal portfolio.
Portfolios In-sample Out-of-sample
Mean Std Dev Sharpe Mean Std Dev Sharpe Average total Average No. of
Ratio Ratio short position assets held long
Panel A: Tangency portfolios
Sample cov matrix, constrained 32.15 15.34 2.10 7.11 20.12 0.35 0.0 17.7
One-factor model 143.08 15.99 8.95 22.14 226.17 0.10 -533.7 177.9
One-factor model, constrained 34.19 15.69 2.18 7.69 21.09 0.36 0.0 27.0
FF 3-factor model 185.99 20.24 9.19 -11.97 128.16 -0.09 -704.3 203.3
FF 3-factor model, constrained 32.76 15.69 2.09 8.17 20.12 0.41 0.0 23.9
Ledoit estimator 264.08 28.63 9.22 3.36 107.28 0.03 -1194.3 217.8
Ledoit estimator, constrained 33.94 16.05 2.11 7.52 20.52 0.37 0.0 24.8
Equally weighted portfolio 7.95 17.54 0.45 0.0 500.0
Panel B: Global minimum variance portfolios
One-factor model 2.80 2.56 1.09 7.42 11.67 0.64 -51.0 268.0
FF 3-factor model 4.41 2.74 1.61 6.46 11.33 0.57 -63.9 283.5
3
6
Table IX
The Ex Post Performance of the Tangency and Global Minimum Variance Portfolios
The assets are the Fama-French 25 size/book-to-market sorted portfolios. The covariance matrix is estimated at the end of April, using
monthly returns of the previous ve years. The tangency portfolios and the global minimum variance portfolios are then formed and
held for one year. This procedure is repeated from 1968 to 1999. The table reports characteristics of ex post excess returns of these
portfolios. Means and standard deviations are those of annual excess returns in excess of the risk-free rate.
Portfolios In-sample Out-of-sample
Mean Std Dev Sharpe Mean Std Dev Sharpe Average total Average No. of
Ratio Ratio short positioin assets held long
Panel A: Tangency portfolios
Sample cov matrix 61.00 17.23 3.54 28.98 82.44 0.35 -1351.6 8.4
One-factor model 91.32 31.29 2.92 61.77 158.63 0.39 -784.8 10.4
FF 3-factor model 298.49 116.03 2.57 -12.87 257.96 -0.05 -3972.4 10.2
Panel B: Global minimum variance portfolios
Sample cov matrix 9.06 7.44 1.22 9.63 15.93 0.60 -481.5 13.0
One-factor model 7.77 8.40 0.92 8.25 16.68 0.49 -147.9 12.4
FF 3-factor model 8.19 8.61 0.95 10.25 14.31 0.72 -252.4 13.0
3
7

Risk Reduction in Large Portfolios: Why Imposing The Wrong Constraints Helps

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Risk Reduction in Large Portfolios: Why Imposing The Wrong Constraints Helps

Uploaded by

Copyright:

Available Formats

Risk Reduction in Large Portfolios:

Why Imposing the Wrong Constraints Helps

are the Lagrange multipliers for the nonnegativity constraints (3), =

the multipliers for the constraints (4), and

). Later we will interpret

be iid normal N(, ). Then the MLE of

/T. The likelihood function depends on both and , even though

), and we can simplify the statements in Proposition

). Notice that the eect of imposing the

, which varies from 0 to 0.4 across dierent sets of simulations.

= 0.4, represent the cases where the nonnegativity

> 0 and the covariance matrix estimate is the sample

= 0.1, the nonnegativity constraint always improves the out-of-

goes from 0.2 to 0.4, we always nd cuto points for

= 0.4 roughly corresponds to the covariance

= 0.2 or 0.3 represents cases where the nonnegativity

= 0.4. Since these cuto points are not sensitive

1)|. Obviously this is always nonnegative. So

S S) -0.143 0.514 527.7

S S) 0.296 0.239 157.6

, the betas with respect

0.1 0.2 0.3 0.4

is the cross-sectional standard

is 0.4, the nonnegativity constraint is violated roughly

is 0.2, the nonnegativity constraint is

is 0.4. The rows with T = Innity are the population results.

, N) pair, we report the

0.2 0.3 0.4

You might also like